Qwen3.6-35B-A3B Local Instance Outperforms Claude Opus 4.7 on SVG Pelican Benchmark
Quantized 35B Qwen model on consumer laptop beats Claude Opus 4.7 at SVG creative benchmark, underscoring open-source gains in local, decentralized AI capabilities.
Simon Willison reported that a 20.9GB quantized Qwen3.6-35B-A3B-UD-Q4_K_S.gguf model running via LM Studio on a MacBook Pro M5 produced a superior SVG of a pelican riding a bicycle than Anthropic's Claude Opus 4.7, which failed on bicycle frame geometry (simonwillison.net, 2026). A follow-up flamingo-on-unicycle SVG test also favored Qwen, including an inline comment noting "Sunglasses on flamingo!" Willison explicitly stated he does not believe Qwen trained on his benchmark.
Primary coverage omitted that SVG generation is a code synthesis task where Qwen's documented strengths in instruction following and structured output excel, per Alibaba's Qwen2 technical report (arXiv:2407.10671, 2024). Related releases such as Meta's Llama 3 Herd of Models (arXiv:2407.21783, 2024) and subsequent open model quantization advances via Unsloth and llama.cpp show consistent gains in local creative coding benchmarks since late 2024, a pattern missed in the original post's focus on the benchmark's absurdity. Claude Opus 4.7's errors align with documented proprietary model inconsistencies on precise vector geometry reported in LMSYS Arena evaluations.
Synthesizing the Simon Willison test, Qwen technical reports, and Hugging Face Open LLM Leaderboard trends demonstrates open-source 30B-scale models now match or exceed frontier systems on narrow creative tasks when quantized to 4-bit and run locally. This reflects accelerating progress toward accessible, decentralized inference without API dependency or vendor lock-in.
AXIOM: Quantized open-source models like Qwen3.6 running locally now surpass select proprietary frontier systems on creative coding tasks, accelerating the shift to decentralized AI infrastructure that reduces reliance on centralized cloud providers.
Sources (3)
- [1]Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7(https://simonwillison.net/2026/Apr/16/qwen-beats-opus/)
- [2]Qwen2 Technical Report(https://arxiv.org/abs/2407.10671)
- [3]The Llama 3 Herd of Models(https://arxiv.org/abs/2407.21783)